Your AI Bill Is Going Up. Here’s Why That Isn’t the Problem You Think It Is.

Prices per token have fallen 98% since 2022. Enterprise AI bills have more than tripled. If that math bothers you, good. It should shape your next budget conversation.

I spend most of my week talking with enterprise leaders about AI, and one topic has taken over every conversation since spring. It usually starts the same way. A CFO looks at an invoice, or a CIO gets a renewal quote, and asks some version of the same question: why is this so much higher than we planned?

Here is the number that reframes everything. The price of a token, the basic unit of AI, has dropped about 98% since late 2022. Over that same stretch, the average enterprise AI budget went from roughly $1.2 million to $7 million a year. Prices collapsed and bills tripled at the same time. That is not a contradiction, and understanding why is the difference between a program you control and one that controls you.

Cheaper never meant smaller

When something useful gets cheap, people use far more of it. We watched this exact movie with cloud a decade ago. Storage and compute got cheaper every year, and cloud bills went up every year, because falling unit prices invited more usage than the savings could offset. AI is running the same play, only faster.

The specific accelerant this time is agents. Through 2024, most enterprise AI was a chatbot: one question, one answer, a few hundred tokens. Cheap and predictable. Then agentic tools arrived. An agent does not just answer. It plans, calls tools, reads the results, checks its own work, and retries when it fails. On every step it re-sends its entire history, so token usage compounds fast, and usually invisibly to whoever owns the budget.

The figures are striking. Agentic workflows use 5 to 30 times the tokens of a simple query. The same task run twice can differ 30-fold in cost, depending on how much the agent has to iterate. In one analysis, the agent’s own review-and-retry loop accounted for 59% of all tokens consumed. The expensive part is not the answer. It is the back-and-forth. And no per-seat budget model can anticipate that, because cost no longer tracks how many people have licenses. It tracks how your solutions are designed.

The invoice is one of nine costs

Here is the trap I see most often. Teams forecast AI cost by looking at the token invoice, because that is the number that actually arrives as a bill. But the token invoice is one of nine cost areas in a real AI program. The other eight, infrastructure and GPUs, integration, data pipelines, monitoring, governance, security and compliance, change management, and the cleanup when things fail, are unmetered and easy to overlook. A forecast built on the invoice alone will understate the truth every time. Running the models, what the industry calls inference, is now roughly 85% of enterprise AI budgets, and at scaling AI-native firms it eats about 23% of revenue.

The good news is that most of it is controllable

If cost followed model prices, you would be stuck waiting on vendors. It does not. It follows design choices you control, which means the levers are in your hands. A few move the needle more than anything else.

Routing is the biggest one. Most organizations send every request, including routine work, to their most capable and most expensive model, simply because that is how the first proof of concept was built. Matching each task to the cheapest model that can handle it well typically cuts 60 to 80% per query. The data backs this up: in early 2026, organizations that routed by task paid a median of $2.31 per million tokens, while those defaulting everything to top-tier models paid $18.40. That is an 87% difference from a single design decision.

The rest compound on top. Caching stops you re-paying full price for the same context and can cut the repeated portion by 70 to 90%. Batch processing runs non-urgent work at roughly half price. Context hygiene, sending the model only what a task needs instead of entire document sets, quietly removes a surprising amount of waste. None of these mean doing less with AI. They mean the cheapest token is the one you never send.

Then there is the structural question of where AI runs. Cloud APIs are the right answer for spiky, unpredictable, or genuinely frontier-grade work. But for steady, high-volume workloads, self-hosting on your own infrastructure starts to make financial sense, and compliance or data-residency requirements sometimes force the issue regardless of price. Most organizations I work with land on a deliberate hybrid: efficient in-house models for routine volume, cloud frontier models for the hard reasoning. Documented outcomes run 30 to 70% total cost reduction at scale. If you last modeled this a year ago, the latest accelerators have changed the math enough that it is worth revisiting.

Put guardrails in before you scale, not after

The organizations that got surprised this year were not reckless, and their tools were not broken. They scaled first and added controls after the invoice landed. The ones that stayed in control did three unglamorous things first. They made spend visible, broken down by team, workload, and model, with an owner attached to every dollar. They set limits, with run-rate alerts during the month rather than a surprise at the end of it. And they tied spend to outcomes, so the number could be defended to the board. The metric that matters is not cost per token. It is value per token: what did each resolved ticket, shipped feature, or closed case actually cost, and was it worth it?

Where this goes

Demand is not slowing. Token consumption is projected to grow more than 20-fold by 2030. At the same time, analysts expect a large share of agent projects to be canceled by 2027, and the reason will be cost and unproven value, not the technology. Both things are true at once. The technology is here to stay and will keep growing, and the programs that win will be the ones that can measure their economics and steer them.

So if your AI bill is climbing, that by itself is not the problem. A rising bill can mean you are getting real value, or it can mean spend is running ahead of outcomes with nobody watching. The only way to know which is to get a clear picture: what you are spending, on what, and to what end. Everything else gets easier from there. If you have not had that look yet, this is the quarter to do it.

The most useful first step is simple: get a clear picture of what you are spending on AI, broken down by team, workload, and model. If you would find it helpful to talk through where your program stands, I am always glad to compare notes.

Figures cited are drawn from widely reported 2026 industry data, including the FinOps Foundation, Gartner, Goldman Sachs, and enterprise spend analyses. Model prices move quarterly; treat specific rates as directional.